SAR Image Small Target Detection Algorithm Based on Improved YOLOv8

To address the problem that small targets in SAR target detection often appear as scattering points with high brightness and insignificant features, and are easily confused with noise or background, we propose a novel algorithm based on improved YOLOv8. Firstly, considering the small targets in SAR images, adding a small target layer into the neck network to capture details. Secondly, the global feature pyramid network is combined to produce better fused features. And a global attention mechanism is introduced to reduce feature loss and amplify features in the global dimension. Finally, the generalized Focal Loss is used to improve the miss-detection false detection in target aggregation scenarios. The experimental results show that the improved algorithm achieves a detection accuracy of 93.1 % on the MSAR dataset, which is 3.2 percentage points higher than the YOLOv8s algorithm, thereby enhancing the algorithm's ability to detect small targets in complex backgrounds.